[gmx-developers] gromacs 2016 beta1 bug somewhere in the new simd for bonded/LJ/lincs

[Carlo Camilloni]

Dear GMX developers,

First of all thanks for your continuos efforts in making gromacs better.

I have done a very basic calculation with the new gromacs-2016-beta1 and I think there is a bug somewhere
in the force calculation. (all calculations are run using 1 processor, no mpi, no openmp, no gpus)

I have run a calculation on alanine dipeptide:

1) GMX507: 2 steps using 1 processor (avx2 simd and nothing else, the tpr is generated with gromacs 507)
2) rerun with GMX512 on the trr file
3) rerun with GMX2016 on the trr file
and then I compared forces, velocities and energies, and while they are essentially the same for the step 0
in the case of GMX507 and 512 they are not the same anymore for GMX2016b1.

The differences are not huge, even if from a rerun I would expect no differences at all, in particular on the
forces, so I have run a metadynamics calculation from the same tpr file and using the three codes
patched with plumed to see the effect on the free energy along the phi torsion angle, and unfortunately the 
difference is quite significant.

to be sure I repeated the same tests on a linux workstation (the first test was on my macbook), so different
compilers and so on, but always avx2 simd)

here you can find everything (all the files I have generated):
https://dl.dropboxusercontent.com/u/1349402/bug-report-aladip-2016.tgz <https://dl.dropboxusercontent.com/u/1349402/bug-report-aladip-2016.tgz>

I hope this helps.

Best,
Carlo




-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://maillist.sys.kth.se/pipermail/gromacs.org_gmx-developers/attachments/20160513/e4605fd4/attachment.html>